Exploiting Diurnal User Mobility for Predicting Cell ... - metis 2020
Exploiting Diurnal User Mobility for Predicting Cell Transitions Nandish P. Kuruvatti, Andreas Klein, J¨org Schneider, Hans D. Schotten Institute for Wireless Communications and Navigation University of Kaiserslautern Email: {kuruvatti,aklein,schneider,schotten}@eit.uni-kl.de Abstract—Mobility of commuters is not purely random but rather direction oriented and may be learned after monitoring user movements for a couple of business days. Exploiting movement data and context information of diurnal user movements (public transportation, vehicular users, etc.) allows for predicting cell transitions and lays the basis e.g. for designing efficient resource reservation schemes or smart resource mapping approaches. In real life scenarios, several mobile users co-travel in public transport forming data intensive moving user clusters or moving networks [12]. Various load balancing solutions exist to manage congestion situations that could arise [4][5]. However, the crucial trigger for these solutions is timely prediction of arrival of moving user clusters or moving networks into a cell. This paper presents prediction and detection schemes that exploit context information for predicting user cell transitions and resulting congestion. These schemes are utilized to anticipate the arrival of data intensive moving user groups/moving networks, which are also referred to as ”hotspots”, into a cell. Simulation results demonstrate robust and timely prediction of these events and their applicability for handover optimization and smart resource management even at high velocities. Index Terms- Moving network, moving user cluster, hotspot, cell transition prediction, diurnal mobility, context information
I.
I NTRODUCTION
The advancement in cellular technologies as well as user equipments (UEs) has led to a drastic growth of mobile subscribers. By the end of 2011 [2], the total number of mobile subscriptions was around 6 billion and is expected to reach approximately 9 billion by 2017. Forecasts of few telecom companies [1][2] indicate that wireless traffic growth will be in order of 1000 times larger in 2020 as opposed to 2010. Further, in day to day scenarios such as public transportation (e.g. buses, trains), where groups of mobile users are traveling together, the data traffic demanded by these users is massive due to growing popularity of mobile multimedia services [1]. In a vehicle with advanced communication and networking capabilities, a mobile router situated within vehicle could be managing all user connections in the vehicle, in turn limiting required UE transmit powers. This constellation is referred to as moving network and may become reality in the not too distant future as stated in [10]. In contrast, a group of users traveling in a conventional vehicle, where connections are individually managed by the serving base station (BS), is referred to as moving user cluster. These variants are expected to travel within service areas of mobile network providers, leading to dynamically changing and potentially high traffic demands. Given the forecasts, data traffic demands of moving networks/user clusters will keep increasing and
will impose challenges on traffic and mobility management. One of the most significant problems caused by data intensive moving user clusters or moving networks is congestion due to ”hotspot” situation in a cell. The hotspot caused by such moving entities can be classified as preferential mobility based hotspot [3]. Several conventional techniques such as mobility load balancing [4], channel borrowing [13], coverage adaptation, etc. [5][14] exist to combat hotspot. Further novel approaches for smart resource mapping in context aware multiRAT scenarios are in their inception [6]. However, the basis for these solutions is prediction and detection of arrival of data intensive moving networks/user groups into a cell. There are several schemes in literature which investigate hotspot situation based on network load, blocking or dropping rates [3][7]. The majority of works considers high user arrival rate, low departure rate, or increased bandwidth demand of existing users leading to hotspot in a cell [3]. Although, these are good theoretical representation of hotspot, it is not beneficial to consider such a model in real world hotspot scenarios caused by moving user clusters or moving networks. The work presented in this paper emphasizes on hotspot problem resulting from moving networks/user groups. An approach based on movement estimation is presented to predict and detect arrival of moving networks/user groups into a cell. User cell transitions are predicted well in advance and this context is beneficially applied for pro-actively triggering load balancing mechanisms as potential countermeasures for combating congestion. Outline. Section II deals with methods to predict cell transitions of moving users, section III discusses simulation scenarios and results, and section IV provides conclusion and indicates future work.
II.
C ELL T RANSITION P REDICTION
Almost all types of public transportation, such as trains, buses, and trams, follow diurnal mobility [8]. Mobility of commuters is not purely random but rather direction oriented characterized by origin and destination points [8]. When such mobility is confined to urban regions, the probability of hotspot occurrence is high in cells through which it traverses. Whenever a mobile hotspot is detected in a cell, it is beneficial to predict the next cell to which this moving user group/moving network would travel. This enables prediction of hotspot situation in near future of neighboring cells.
Method 0
A. Diurnal Mobility Model In case of random walk mobility model a user can travel in all six directions with equal probability from its current cell. However in diurnal mobility model (as depicted in Fig. 1), 2
4
0
v
User direction
Method 2
2
6
Fig. 2.
Direction estimation methods
5
3
Fig. 1.
Method 3
v
1
1
Method 1
1) 2)
Diurnal mobility model
user group direction is probable in only two directions (e.g. streets, train tracks) and zero in other directions [8]. Hence, a user group can transit into one of the two adjacent cells from its present cell. If φ1 and φ2 are the angles of users’ direction with respect to the closest directions leading towards center of neighboring cells, then probability of user transition into those neighbors are [8], φ1 p1 = 1− (1) 60 p2
=
φ2 1− 60
3) 4)
The distance of user from cell center is monitored at fixed logging intervals. IF the distance is greater than radius of circle: GOTO step 3, ELSE: user is in the circle, STOP. IF present distance is greater than previous distance: user will not enter the circle, ELSE: repeat step 2. Instantaneous user angle is used to estimate direction. User trajectory Distance from center Shortest distance to center Center of cell
(2)
B. Estimation of User Direction In this section the different approaches for sampling user positions and for estimating user directions are described. The model employs a ”virtual” circle inscribed in each cell. This ”virtual” circle corresponds to a certain signal strength threshold derived from radio propagation data. The center of circle coincides with center of cell. The position of user is estimated at fix, but velocity-dependent intervals. The user positions falling in the circle are recorded and used to estimate direction. The user angle is calculated at each position as, y2 − y1 φ = tan− 1 (3) x2 − x1 where (x2 ,y2 ) and (x1 ,y1 ) are present and previous positions, respectively. Different methods for estimating user direction are depicted in Fig. 2. Method 0 utilizes average of all user angles recorded in the circle to estimate user direction, whereas method 1 only considers average of user angles recorded within the circular strip. Method 2 performs exponential moving average (EMA) [11] on the user angles recorded in the circle and method 3 uses only instantaneous angle of user before leaving the ”virtual” circle. In certain special cases users will not enter the circular region, as depicted in Fig. 3. In order to estimate user direction in such cases, the following algorithm is employed:
Fig. 3.
User not entering recording circle
C. Prediction of Cell Transition Based on Angular Deviation The user abiding to diurnal mobility can transit from its present cell to one of the two neighboring cells [8] influenced by its direction. Table. I lists potential next cells for such user in cell 0, as shown in Fig. 1, based on user angle. Similarly, next cells can be stated for each cell in the cellular layout. The probability of transition to next cells is given by Eq. 1 and Eq. 2. TABLE I. User angle φ 0-60 60-120 120-180 180-240 240-300 300-360
N EXT CELL BASED ON USER ANGLE Set of Next Cells 6,4 4,2 2,1 1,3 3,5 5,6
φ1 φ φ − 60 φ − 120 φ − 180 φ − 240 φ − 300
φ2 60 − φ 120 − φ 180 − φ 240 − φ 300 − φ 360 − φ
D. Comparison of Direction Estimation Methods At high velocity and straight motion of users, all methods mentioned in Fig. 2 yield same prediction results. When user
Cell 4
Cell 2
trajectory deviates within angular range for same next cells (e.g. 0 − 60◦ ), all approaches predict same set of next cells, although estimated directions are different. This is illustrated by trajectory A in Fig. 4 and the results at two different velocities (20 and 100 km/h) are listed in Table II. The actual next cell is 6.
d2 d2 d1
Average EMA Circular strip Instantaneous
d1
60-120 deg
Cell 0
0-60 deg
Fig. 5.
Cell 1
Cell transition prediction based on distance
Cell 0
if d1 and d2 are the distances of user from centers of cell 1 and cell 2 then probabilities of transition to these cells based on distance are,
trajectory A trajectory B
Fig. 4.
Illustration of predicted directions for various user trajectories
In case the user trajectory deviates outside angular range for same next cells, instantaneous and circular strip-based approaches lead to prediction of actual next cells. EMA and average lead to prediction of different set of next cells, due to consideration of history values of angles. Such a scenario is depicted by trajectory B in Fig. 4 and the results at two different velocities (20 and 100 km/h) are listed in Table III, actual next cell being cell 4. For high velocities, sampling rate is to be adapted in order to yield robust estimation in particular for history-based approaches. TABLE II.
C OMPARISON OF E STIMATION M ETHODS (W ITHIN R ANGE )
Estimation Method Average EMA Circular Strip Instantaneous Average EMA Circular Strip Instantaneous
TABLE III. Estimation Method Average EMA Circular Strip Instantaneous Average EMA Circular Strip Instantaneous
Velocity (km/h) 20
100
Angular Deviation 24.12 23.39 18 18 24.6 21.4 18 18
Set of Next Cells 6,4 6,4 6,4 6,4 6,4 6,4 6,4 6,4
P1
P2
0.598 0.611 0.7 0.7 0.59 0.643 0.7 0.7
0.402 0.389 0.3 0.3 0.41 0.357 0.3 0.3
C OMPARISON OF E STIMATION M ETHODS (O UTSIDE R ANGE ) Velocity (km/h) 20
100
Angular Deviation 55 56 63 65 55 59 65 65
Set of Next Cells 6,4 6,4 4,2 4,2 6,4 6,4 4,2 4,2
P1
P2
0.084 0.067 0.95 0.917 0.084 0.017 0.917 0.917
0.916 0.933 0.05 0.083 0.916 0.983 0.083 0.083
E. Prediction of Cell Transition Based on Distance The user movements indicated in green and blue lines in Fig. 5 have the same estimated user direction (angle). But the next cell of transition depends on position of user trajectory at the circumference of the circle. At the point of prediction,
p1
=
p2
=
d1 , d1 + d2 d2 , 1− d1 + d2
1−
(4) (5)
F. Combined approach Probabilities of transition are different for angle-based and distance-based methods. The probabilities can be combined by taking average of the two. However, distance based approach is observed to have more impact on probability than method employing user angle. Hence, the distance component of probability equation is weighed by α >1. The probability equations of the combined approach are, α d1 φ1 − · , 60(1 + α) 1 + α d1 + d2 α d2 φ2 − · p2 = 1 − , 60(1 + α) 1 + α d1 + d2
p1 = 1 −
(6) (7)
G. Special Case In Fig. 6, user is traveling from cell 3 following diurnal mobility. In normal case user always transits to one of the two next cells based on its direction [8]. In special case, a brief transition (indicated in red) happens to a third cell before moving to probable next cell. For instance, in cell 6 it is estimated that cell 15 and cell 18 are the next cells when relying on angle-based method. But user briefly travels to cell 4 which cannot be traced. The brief transition to a third cell could be predicted by considering three potential next cells instead of two for each user direction range. The probabilities of transition in these cells based on angle [8] are, φ1 p1 = 1− (8) 60 φ2 p2 = 1− (9) 60 p3 = 0 (10)
TABLE V. 49
47 28
26 11 2
15
4
3
80 53
Present Cell
113 115
Cell 3
82
32
Cell 0
18
6 5
14
51 30
13
0
1
78
17
Cell 6
16
Cell 4
Fig. 6.
N EXT C ELL P REDICTION R ESULTS
Next Cells Cell 5 Cell 0 Cell 1 Cell 6 Cell 4 Cell 2 Cell 18 Cell 15 Cell 4 Cell 15 Cell 13 Cell 11
Pangle 0.25 0.75 0 0.25 0.75 0 0.25 0.75 0 0.25 0.75 0
Pdistance 0.346 0.410 0.243 0.447 0.353 0.198 0.202 0.332 0.465 0.468 0.311 0.219
Pcombined 0.327 0.478 0.194 0.408 0.433 0.158 0.211 0.416 0.372 0.424 0.399 0.175
User trajectory in special case
At prediction point, if d1 , d2 , and d3 are distances of user from centers of next cells then probabilities based on distance are, p1
=
p2
=
p3
=
d1 , d1 + d2 + d3 d2 1− , d1 + d2 + d3 d3 , 1− d1 + d2 + d3 1−
(11) (12) (13)
Combined probabilities similar to section II-F are given by,
From the table, it could be observed that the angle-based approach can’t trace transition into a third cell and its probability is always zero. This limitation is also reflected in combined approach, affecting prediction accuracy even with high values of α. The affect of α on prediction of next cell during brief transition (special case) is described in Fig. 7. With increase in value of α, prediction in actual next cell by combined approach improves but does not exceed the probability of prediction in other cell (e.g. next cell 2 in Fig. 7). Fig. 8 illustrates the comparison among distance based, angle based and combined approaches in predicting transition into actual cell. It could be seen that distance based approach makes best prediction with consistency among the three. Affect of α on Prediction Result
3 + 2α φ1 α d1 p1 = − − · , (14) 3(1 + α) 60(1 + α) 1 + α d1 + d2 + d3 φ2 α d2 3 + 2α − − · , (15) p2 = 3(1 + α) 60(1 + α) 1 + α d1 + d2 + d3 2α α d3 p3 = − · , (16) 3(1 + α) 1 + α d1 + d2 + d3
Probability of Prediction
S IMULATION R ESULTS
A system-level simulator is used and a multi-cell scenario is created as illustrated in Fig. 6 with a BS at each cell center. The considered radio access technology is LTE. The cluster of 60 users moving together at high velocity (120 km/h) constitutes a mobile hotspot. Each cell is allowed to have static background users. Evaluation methodology follows [9] and assumes 10 MHz bandwidth for LTE operation at 2 GHz. The trajectory followed by moving user group is as in Fig. 6. Table IV summarizes simulation parameters. TABLE IV. Parameter Carrier frequency System bandwidth Total transmit power Control channel overhead Shadowing Fast fading Noise power Background users per cell Hotspot users
Table V shows results of next cell prediction for user trajectory in Fig. 6 from cell 3 to cell 15.
0.45 0.4 0.35 0.3 0.25 0.2 1
Fig. 7.
2
3 α
4
5
Affect of α on Prediction Probability Prediction of Actual Cells 0.8 angle based distance based combination
0.7
S IMULATION PARAMETERS Assumption 2 GHz 10 MHz (50 PRBs) 40 W/250 mW (s2s = 500m/200m) 12% log-normal Standard deviation: 8 dB Decorrelation distance: 50 m 2-tap Rayleigh fading channel −174 dBm/Hz + 10 · log10 (B) + 7 30 60 at 120 km/h
Next cell 1 Next cell 2 Actual next cell
0.5
Probability of Prediction
III.
0.55
0.6 0.5 0.4 0.3 0.2 0.1 0
Fig. 8.
Prediction Events
Comparison of Prediction Approaches
The earliness of next cell prediction could be improved by reducing radius of recording circle. However this has a tradeoff with accuracy of prediction. Further, prediction at
low velocity is earlier than at high velocity since there is a longer period of time (dwelling period) between prediction and actual transition at lower velocities.In general, prediction of hotspot and next cell transition are easier at low velocities and becomes challenging at high velocities. The simulations are carried out for site-to-site distances (s2s) of 500m and 200m (dense deployment), respectively. The moving user group travels same distance in both cases. Load balancing (LB) mechanism is triggered in respective cells which anticipate arrival of moving networks/user groups enabled by prediction of user cell transitions. This procedure relieves congestion and accommodates for approaching moving networks/user groups. This is reflected by reduced dropping of connections, reduction in blocked access attempts, and decrease in blocked handovers at respective BSs. There is only a very slight decrease ( 1%) in load, since the freed up resources will soon be occupied by moving networks/user groups. The average reduction in these KPIs using LB triggered by cell transition prediction for s2s = 500m and s2s = 200m, respectively, are illustrated in Fig. 9 and Fig. 10. These results demonstrate that prediction
monitoring user movements for a couple of business days, is exploited to predict cell transitions. This context information is utilized to trigger load balancing as a potential countermeasure to combat eminent congestion. Simulation results demonstrate the ability of presented approaches to predict arrival of moving user groups/networks well in advance, maintaining consistency even at high velocities. Future work aims at applying these prediction approaches for designing smart radio resource management and load balancing strategies in heterogeneous and ultra-dense deployments. V.
Part of this work has been performed in the framework of FP7 project ICT-317669 METIS, which is partly funded by the European Union. The authors alone are responsible for the content of the paper. R EFERENCES [1]
Dropping 1
w/o pred. w. pred.+LB
Average (%)
0.8 0.6
[2] [3]
0.4 0.2 0
s2s=500
[4]
s2s=200
Blocked Handover attempts 1
w/o pred. w. pred.+LB
Average (%)
0.8
[5]
0.6 0.4 0.2 0
Fig. 9.
[6] s2s=500
s2s=200
Dropping and blocked HO attempts
[7]
Blocked Access attempts 1 w/o pred. w. pred.+LB
Average (%)
0.8
[8]
0.6 0.4 0.2 0
s2s=500
[9]
s2s=200 Load
1 w/o pred. w. pred.+LB
Average (%)
0.8
[11]
0.2 0
Fig. 10.
[10]
0.6 0.4
s2s=500
s2s=200
Blocked access attempts and load
of user cell transitions could be exploited to anticipate arrival of moving networks/user groups into a cell and, thus laying a basis for context-aware radio resource management. The results show the validity of these schemes even at high velocities as well as dense deployments, which are envisioned to spread in future [1][10]. IV.
C ONCLUSION AND F UTURE W ORK
The primary step to mitigate congestion issues caused by moving user groups/networks is to predict their arrival into a cell. This paper presented and evaluated methods for hotspot prediction with an emphasis on moving user clusters/networks. The diurnal mobility of users, that can be learned after
ACKNOWLEDGMENT
[12]
[13]
[14]
Nokia Siemens Networks, 2020: Beyond 4G Radio Evolution for the Gigabit Experience, white paper, Aug., 2011. Ericsson, More than 50 billion connected devices, white paper, Feb., 2011. J. Jobin, M. Faloutsos, S. K. Tripathi and S.V. Krishnamurthy, Understanding the Effects of Hotspots in Wireless Cellular Networks, INFOCOM 2004. A. Lobinger, S. Stefanski, T. Jansen, and I. Balan, Load Balancing in Downlink LTE Self-Optimizing Networks, 71st IEEE Vehicular Technology Conference (VTC-Spring), May, 2010. A. Sharma, A. Roy, S. Ghosal, R. Chaki and U. Bhattacharya, Load balancing in Cellular Network: A review, International Conference on Computing, Communication & Networking Technologies (ICCNT) 2012. A. Klein, C. Mannweiler, J. Schneider, F. Thillen and H.D. Schotten, A concept for context-enhanced heterogeneous access management, IEEE Globecom workshops 2010. Y.K Tang, J. Misic, H. Zhu and I. Chlamtac, An Online Hotspot Detection Scheme in DS-CDMA Networks - Single Traffic Type, IEEE Globecom 2004. S.K. Sadhukhan, S. Mandal, P. Bhaumik and D. Saha, A Novel Direction-Based Diurnal Mobility Model for Handoff Estimation in Cellular Networks, IEEE India conference (INDICON) 2010. 3GPP TSG RAN, Further advancements for E-UTRA physical layer aspects (Release 9), TR 36.814 V9.0.0, March, 2010. METIS, D1.1 - Future radio access scenarios, requirements and KPIs, May, 2013, available online: https://www.metis2020.com. D. Haynes, J. Misic, S. Corns and G.K Venayagamoorthy, An Exponential Moving Average algorithm, IEEE Congress on Evolutionary Computation 2012. J.l Agbinya, Design concepts: wireless moving networks (WMN), The 6th International Conference on Advanced Communication Technology, 2004. Quan Lin, L. Nahar and S. Bhattacharya, A remote channel borrowing approach for real-time congestion control in wireless communication, IEEE Globecom 1997. C. Askarian and H. Beigy A Survey for Load Balancing in Mobile WiMAX Networks, Advanced Computing: An International Journal ( ACIJ ), Vol.3, No.2, March 2012.
I.
I NTRODUCTION
The advancement in cellular technologies as well as user equipments (UEs) has led to a drastic growth of mobile subscribers. By the end of 2011 [2], the total number of mobile subscriptions was around 6 billion and is expected to reach approximately 9 billion by 2017. Forecasts of few telecom companies [1][2] indicate that wireless traffic growth will be in order of 1000 times larger in 2020 as opposed to 2010. Further, in day to day scenarios such as public transportation (e.g. buses, trains), where groups of mobile users are traveling together, the data traffic demanded by these users is massive due to growing popularity of mobile multimedia services [1]. In a vehicle with advanced communication and networking capabilities, a mobile router situated within vehicle could be managing all user connections in the vehicle, in turn limiting required UE transmit powers. This constellation is referred to as moving network and may become reality in the not too distant future as stated in [10]. In contrast, a group of users traveling in a conventional vehicle, where connections are individually managed by the serving base station (BS), is referred to as moving user cluster. These variants are expected to travel within service areas of mobile network providers, leading to dynamically changing and potentially high traffic demands. Given the forecasts, data traffic demands of moving networks/user clusters will keep increasing and
will impose challenges on traffic and mobility management. One of the most significant problems caused by data intensive moving user clusters or moving networks is congestion due to ”hotspot” situation in a cell. The hotspot caused by such moving entities can be classified as preferential mobility based hotspot [3]. Several conventional techniques such as mobility load balancing [4], channel borrowing [13], coverage adaptation, etc. [5][14] exist to combat hotspot. Further novel approaches for smart resource mapping in context aware multiRAT scenarios are in their inception [6]. However, the basis for these solutions is prediction and detection of arrival of data intensive moving networks/user groups into a cell. There are several schemes in literature which investigate hotspot situation based on network load, blocking or dropping rates [3][7]. The majority of works considers high user arrival rate, low departure rate, or increased bandwidth demand of existing users leading to hotspot in a cell [3]. Although, these are good theoretical representation of hotspot, it is not beneficial to consider such a model in real world hotspot scenarios caused by moving user clusters or moving networks. The work presented in this paper emphasizes on hotspot problem resulting from moving networks/user groups. An approach based on movement estimation is presented to predict and detect arrival of moving networks/user groups into a cell. User cell transitions are predicted well in advance and this context is beneficially applied for pro-actively triggering load balancing mechanisms as potential countermeasures for combating congestion. Outline. Section II deals with methods to predict cell transitions of moving users, section III discusses simulation scenarios and results, and section IV provides conclusion and indicates future work.
II.
C ELL T RANSITION P REDICTION
Almost all types of public transportation, such as trains, buses, and trams, follow diurnal mobility [8]. Mobility of commuters is not purely random but rather direction oriented characterized by origin and destination points [8]. When such mobility is confined to urban regions, the probability of hotspot occurrence is high in cells through which it traverses. Whenever a mobile hotspot is detected in a cell, it is beneficial to predict the next cell to which this moving user group/moving network would travel. This enables prediction of hotspot situation in near future of neighboring cells.
Method 0
A. Diurnal Mobility Model In case of random walk mobility model a user can travel in all six directions with equal probability from its current cell. However in diurnal mobility model (as depicted in Fig. 1), 2
4
0
v
User direction
Method 2
2
6
Fig. 2.
Direction estimation methods
5
3
Fig. 1.
Method 3
v
1
1
Method 1
1) 2)
Diurnal mobility model
user group direction is probable in only two directions (e.g. streets, train tracks) and zero in other directions [8]. Hence, a user group can transit into one of the two adjacent cells from its present cell. If φ1 and φ2 are the angles of users’ direction with respect to the closest directions leading towards center of neighboring cells, then probability of user transition into those neighbors are [8], φ1 p1 = 1− (1) 60 p2
=
φ2 1− 60
3) 4)
The distance of user from cell center is monitored at fixed logging intervals. IF the distance is greater than radius of circle: GOTO step 3, ELSE: user is in the circle, STOP. IF present distance is greater than previous distance: user will not enter the circle, ELSE: repeat step 2. Instantaneous user angle is used to estimate direction. User trajectory Distance from center Shortest distance to center Center of cell
(2)
B. Estimation of User Direction In this section the different approaches for sampling user positions and for estimating user directions are described. The model employs a ”virtual” circle inscribed in each cell. This ”virtual” circle corresponds to a certain signal strength threshold derived from radio propagation data. The center of circle coincides with center of cell. The position of user is estimated at fix, but velocity-dependent intervals. The user positions falling in the circle are recorded and used to estimate direction. The user angle is calculated at each position as, y2 − y1 φ = tan− 1 (3) x2 − x1 where (x2 ,y2 ) and (x1 ,y1 ) are present and previous positions, respectively. Different methods for estimating user direction are depicted in Fig. 2. Method 0 utilizes average of all user angles recorded in the circle to estimate user direction, whereas method 1 only considers average of user angles recorded within the circular strip. Method 2 performs exponential moving average (EMA) [11] on the user angles recorded in the circle and method 3 uses only instantaneous angle of user before leaving the ”virtual” circle. In certain special cases users will not enter the circular region, as depicted in Fig. 3. In order to estimate user direction in such cases, the following algorithm is employed:
Fig. 3.
User not entering recording circle
C. Prediction of Cell Transition Based on Angular Deviation The user abiding to diurnal mobility can transit from its present cell to one of the two neighboring cells [8] influenced by its direction. Table. I lists potential next cells for such user in cell 0, as shown in Fig. 1, based on user angle. Similarly, next cells can be stated for each cell in the cellular layout. The probability of transition to next cells is given by Eq. 1 and Eq. 2. TABLE I. User angle φ 0-60 60-120 120-180 180-240 240-300 300-360
N EXT CELL BASED ON USER ANGLE Set of Next Cells 6,4 4,2 2,1 1,3 3,5 5,6
φ1 φ φ − 60 φ − 120 φ − 180 φ − 240 φ − 300
φ2 60 − φ 120 − φ 180 − φ 240 − φ 300 − φ 360 − φ
D. Comparison of Direction Estimation Methods At high velocity and straight motion of users, all methods mentioned in Fig. 2 yield same prediction results. When user
Cell 4
Cell 2
trajectory deviates within angular range for same next cells (e.g. 0 − 60◦ ), all approaches predict same set of next cells, although estimated directions are different. This is illustrated by trajectory A in Fig. 4 and the results at two different velocities (20 and 100 km/h) are listed in Table II. The actual next cell is 6.
d2 d2 d1
Average EMA Circular strip Instantaneous
d1
60-120 deg
Cell 0
0-60 deg
Fig. 5.
Cell 1
Cell transition prediction based on distance
Cell 0
if d1 and d2 are the distances of user from centers of cell 1 and cell 2 then probabilities of transition to these cells based on distance are,
trajectory A trajectory B
Fig. 4.
Illustration of predicted directions for various user trajectories
In case the user trajectory deviates outside angular range for same next cells, instantaneous and circular strip-based approaches lead to prediction of actual next cells. EMA and average lead to prediction of different set of next cells, due to consideration of history values of angles. Such a scenario is depicted by trajectory B in Fig. 4 and the results at two different velocities (20 and 100 km/h) are listed in Table III, actual next cell being cell 4. For high velocities, sampling rate is to be adapted in order to yield robust estimation in particular for history-based approaches. TABLE II.
C OMPARISON OF E STIMATION M ETHODS (W ITHIN R ANGE )
Estimation Method Average EMA Circular Strip Instantaneous Average EMA Circular Strip Instantaneous
TABLE III. Estimation Method Average EMA Circular Strip Instantaneous Average EMA Circular Strip Instantaneous
Velocity (km/h) 20
100
Angular Deviation 24.12 23.39 18 18 24.6 21.4 18 18
Set of Next Cells 6,4 6,4 6,4 6,4 6,4 6,4 6,4 6,4
P1
P2
0.598 0.611 0.7 0.7 0.59 0.643 0.7 0.7
0.402 0.389 0.3 0.3 0.41 0.357 0.3 0.3
C OMPARISON OF E STIMATION M ETHODS (O UTSIDE R ANGE ) Velocity (km/h) 20
100
Angular Deviation 55 56 63 65 55 59 65 65
Set of Next Cells 6,4 6,4 4,2 4,2 6,4 6,4 4,2 4,2
P1
P2
0.084 0.067 0.95 0.917 0.084 0.017 0.917 0.917
0.916 0.933 0.05 0.083 0.916 0.983 0.083 0.083
E. Prediction of Cell Transition Based on Distance The user movements indicated in green and blue lines in Fig. 5 have the same estimated user direction (angle). But the next cell of transition depends on position of user trajectory at the circumference of the circle. At the point of prediction,
p1
=
p2
=
d1 , d1 + d2 d2 , 1− d1 + d2
1−
(4) (5)
F. Combined approach Probabilities of transition are different for angle-based and distance-based methods. The probabilities can be combined by taking average of the two. However, distance based approach is observed to have more impact on probability than method employing user angle. Hence, the distance component of probability equation is weighed by α >1. The probability equations of the combined approach are, α d1 φ1 − · , 60(1 + α) 1 + α d1 + d2 α d2 φ2 − · p2 = 1 − , 60(1 + α) 1 + α d1 + d2
p1 = 1 −
(6) (7)
G. Special Case In Fig. 6, user is traveling from cell 3 following diurnal mobility. In normal case user always transits to one of the two next cells based on its direction [8]. In special case, a brief transition (indicated in red) happens to a third cell before moving to probable next cell. For instance, in cell 6 it is estimated that cell 15 and cell 18 are the next cells when relying on angle-based method. But user briefly travels to cell 4 which cannot be traced. The brief transition to a third cell could be predicted by considering three potential next cells instead of two for each user direction range. The probabilities of transition in these cells based on angle [8] are, φ1 p1 = 1− (8) 60 φ2 p2 = 1− (9) 60 p3 = 0 (10)
TABLE V. 49
47 28
26 11 2
15
4
3
80 53
Present Cell
113 115
Cell 3
82
32
Cell 0
18
6 5
14
51 30
13
0
1
78
17
Cell 6
16
Cell 4
Fig. 6.
N EXT C ELL P REDICTION R ESULTS
Next Cells Cell 5 Cell 0 Cell 1 Cell 6 Cell 4 Cell 2 Cell 18 Cell 15 Cell 4 Cell 15 Cell 13 Cell 11
Pangle 0.25 0.75 0 0.25 0.75 0 0.25 0.75 0 0.25 0.75 0
Pdistance 0.346 0.410 0.243 0.447 0.353 0.198 0.202 0.332 0.465 0.468 0.311 0.219
Pcombined 0.327 0.478 0.194 0.408 0.433 0.158 0.211 0.416 0.372 0.424 0.399 0.175
User trajectory in special case
At prediction point, if d1 , d2 , and d3 are distances of user from centers of next cells then probabilities based on distance are, p1
=
p2
=
p3
=
d1 , d1 + d2 + d3 d2 1− , d1 + d2 + d3 d3 , 1− d1 + d2 + d3 1−
(11) (12) (13)
Combined probabilities similar to section II-F are given by,
From the table, it could be observed that the angle-based approach can’t trace transition into a third cell and its probability is always zero. This limitation is also reflected in combined approach, affecting prediction accuracy even with high values of α. The affect of α on prediction of next cell during brief transition (special case) is described in Fig. 7. With increase in value of α, prediction in actual next cell by combined approach improves but does not exceed the probability of prediction in other cell (e.g. next cell 2 in Fig. 7). Fig. 8 illustrates the comparison among distance based, angle based and combined approaches in predicting transition into actual cell. It could be seen that distance based approach makes best prediction with consistency among the three. Affect of α on Prediction Result
3 + 2α φ1 α d1 p1 = − − · , (14) 3(1 + α) 60(1 + α) 1 + α d1 + d2 + d3 φ2 α d2 3 + 2α − − · , (15) p2 = 3(1 + α) 60(1 + α) 1 + α d1 + d2 + d3 2α α d3 p3 = − · , (16) 3(1 + α) 1 + α d1 + d2 + d3
Probability of Prediction
S IMULATION R ESULTS
A system-level simulator is used and a multi-cell scenario is created as illustrated in Fig. 6 with a BS at each cell center. The considered radio access technology is LTE. The cluster of 60 users moving together at high velocity (120 km/h) constitutes a mobile hotspot. Each cell is allowed to have static background users. Evaluation methodology follows [9] and assumes 10 MHz bandwidth for LTE operation at 2 GHz. The trajectory followed by moving user group is as in Fig. 6. Table IV summarizes simulation parameters. TABLE IV. Parameter Carrier frequency System bandwidth Total transmit power Control channel overhead Shadowing Fast fading Noise power Background users per cell Hotspot users
Table V shows results of next cell prediction for user trajectory in Fig. 6 from cell 3 to cell 15.
0.45 0.4 0.35 0.3 0.25 0.2 1
Fig. 7.
2
3 α
4
5
Affect of α on Prediction Probability Prediction of Actual Cells 0.8 angle based distance based combination
0.7
S IMULATION PARAMETERS Assumption 2 GHz 10 MHz (50 PRBs) 40 W/250 mW (s2s = 500m/200m) 12% log-normal Standard deviation: 8 dB Decorrelation distance: 50 m 2-tap Rayleigh fading channel −174 dBm/Hz + 10 · log10 (B) + 7 30 60 at 120 km/h
Next cell 1 Next cell 2 Actual next cell
0.5
Probability of Prediction
III.
0.55
0.6 0.5 0.4 0.3 0.2 0.1 0
Fig. 8.
Prediction Events
Comparison of Prediction Approaches
The earliness of next cell prediction could be improved by reducing radius of recording circle. However this has a tradeoff with accuracy of prediction. Further, prediction at
low velocity is earlier than at high velocity since there is a longer period of time (dwelling period) between prediction and actual transition at lower velocities.In general, prediction of hotspot and next cell transition are easier at low velocities and becomes challenging at high velocities. The simulations are carried out for site-to-site distances (s2s) of 500m and 200m (dense deployment), respectively. The moving user group travels same distance in both cases. Load balancing (LB) mechanism is triggered in respective cells which anticipate arrival of moving networks/user groups enabled by prediction of user cell transitions. This procedure relieves congestion and accommodates for approaching moving networks/user groups. This is reflected by reduced dropping of connections, reduction in blocked access attempts, and decrease in blocked handovers at respective BSs. There is only a very slight decrease ( 1%) in load, since the freed up resources will soon be occupied by moving networks/user groups. The average reduction in these KPIs using LB triggered by cell transition prediction for s2s = 500m and s2s = 200m, respectively, are illustrated in Fig. 9 and Fig. 10. These results demonstrate that prediction
monitoring user movements for a couple of business days, is exploited to predict cell transitions. This context information is utilized to trigger load balancing as a potential countermeasure to combat eminent congestion. Simulation results demonstrate the ability of presented approaches to predict arrival of moving user groups/networks well in advance, maintaining consistency even at high velocities. Future work aims at applying these prediction approaches for designing smart radio resource management and load balancing strategies in heterogeneous and ultra-dense deployments. V.
Part of this work has been performed in the framework of FP7 project ICT-317669 METIS, which is partly funded by the European Union. The authors alone are responsible for the content of the paper. R EFERENCES [1]
Dropping 1
w/o pred. w. pred.+LB
Average (%)
0.8 0.6
[2] [3]
0.4 0.2 0
s2s=500
[4]
s2s=200
Blocked Handover attempts 1
w/o pred. w. pred.+LB
Average (%)
0.8
[5]
0.6 0.4 0.2 0
Fig. 9.
[6] s2s=500
s2s=200
Dropping and blocked HO attempts
[7]
Blocked Access attempts 1 w/o pred. w. pred.+LB
Average (%)
0.8
[8]
0.6 0.4 0.2 0
s2s=500
[9]
s2s=200 Load
1 w/o pred. w. pred.+LB
Average (%)
0.8
[11]
0.2 0
Fig. 10.
[10]
0.6 0.4
s2s=500
s2s=200
Blocked access attempts and load
of user cell transitions could be exploited to anticipate arrival of moving networks/user groups into a cell and, thus laying a basis for context-aware radio resource management. The results show the validity of these schemes even at high velocities as well as dense deployments, which are envisioned to spread in future [1][10]. IV.
C ONCLUSION AND F UTURE W ORK
The primary step to mitigate congestion issues caused by moving user groups/networks is to predict their arrival into a cell. This paper presented and evaluated methods for hotspot prediction with an emphasis on moving user clusters/networks. The diurnal mobility of users, that can be learned after
ACKNOWLEDGMENT
[12]
[13]
[14]
Nokia Siemens Networks, 2020: Beyond 4G Radio Evolution for the Gigabit Experience, white paper, Aug., 2011. Ericsson, More than 50 billion connected devices, white paper, Feb., 2011. J. Jobin, M. Faloutsos, S. K. Tripathi and S.V. Krishnamurthy, Understanding the Effects of Hotspots in Wireless Cellular Networks, INFOCOM 2004. A. Lobinger, S. Stefanski, T. Jansen, and I. Balan, Load Balancing in Downlink LTE Self-Optimizing Networks, 71st IEEE Vehicular Technology Conference (VTC-Spring), May, 2010. A. Sharma, A. Roy, S. Ghosal, R. Chaki and U. Bhattacharya, Load balancing in Cellular Network: A review, International Conference on Computing, Communication & Networking Technologies (ICCNT) 2012. A. Klein, C. Mannweiler, J. Schneider, F. Thillen and H.D. Schotten, A concept for context-enhanced heterogeneous access management, IEEE Globecom workshops 2010. Y.K Tang, J. Misic, H. Zhu and I. Chlamtac, An Online Hotspot Detection Scheme in DS-CDMA Networks - Single Traffic Type, IEEE Globecom 2004. S.K. Sadhukhan, S. Mandal, P. Bhaumik and D. Saha, A Novel Direction-Based Diurnal Mobility Model for Handoff Estimation in Cellular Networks, IEEE India conference (INDICON) 2010. 3GPP TSG RAN, Further advancements for E-UTRA physical layer aspects (Release 9), TR 36.814 V9.0.0, March, 2010. METIS, D1.1 - Future radio access scenarios, requirements and KPIs, May, 2013, available online: https://www.metis2020.com. D. Haynes, J. Misic, S. Corns and G.K Venayagamoorthy, An Exponential Moving Average algorithm, IEEE Congress on Evolutionary Computation 2012. J.l Agbinya, Design concepts: wireless moving networks (WMN), The 6th International Conference on Advanced Communication Technology, 2004. Quan Lin, L. Nahar and S. Bhattacharya, A remote channel borrowing approach for real-time congestion control in wireless communication, IEEE Globecom 1997. C. Askarian and H. Beigy A Survey for Load Balancing in Mobile WiMAX Networks, Advanced Computing: An International Journal ( ACIJ ), Vol.3, No.2, March 2012.
